Guest Editorial: Special Issue on Grid Computing Jon Weissman Dept. of Computer Science & Engineering University of Minnesota, Minneapolis, MN 55455 ion@cs.umn.edu Computational Grids have been advanced as a platform for solving problems in science and engineering requiring the coupling of distributed, often heterogeneous, resources. To exploit the full potential of Grids, the performance of Grid resources and the applications that run on them, must be modeled to ensure levels of performance that meet end-user expectation. Different performance metrics may be applicable such as predictable or consistent performance, high performance in terms of throughput or turn-around time, or response time or latency. This issue of Sigmetrics PER addresses recent work in the exciting area of Grid Performance. The paper by Valerie Taylor, Xinfu Wu, and Rick Stevens describes Prophesy, a performance analysis and modeling infrastructure for Grid applications. Prophesy provides an environment for archival of data relating to application, executable, run-time, and performance statistics information as well as model development. Prophesy captures the context in which the performance data was collected as well as the data. Key features of Prophesy include generated instrumentation, an automated modeling component that generates a performance model for each application kernel based on techniques such as curve fitting. It also introduces the notion of performance coupling to allow the different kernel models to be combined into a larger model. The paper by Bruce Lowekamp describes the design of a new network-monitoring tool called Wren. It is designed to provide accurate, non-intrnsive network measurements for large-scale networks such as Grids. It combines passive measurement and active probes to form a picture of the network. The novel idea of topology-based steering is introduced to limit the active probes to only bottleneck links thus providing greater scalability and reduced intrusiveness. The paper by Allan Snavely, Greg Chun, Henri Casanova, Rob Van der Wijngaart and Michael Frumkin describe current benchmark strategies and activities for Grid applications. Benchmarking in the Grid is challenging for several reasons: many Grid applications rely on middleware and services, reproducible performance may be difficult to achieve, dedicated use of the Grid for benchmarking may be infeasible, and Grid applications are extremely diverse. The authors participate in the Grid Benchmarking Research Group of the Global Grid Forum and present current work including a Grid application taxonomy and separate benchmarks for compute-intensive and dataintensive codes. For compute-intensive codes, the well-known NAS parallel benchmarks have been extended to the Grid, and early work in data-intensive benchmarks has focused on a fimher sub-classification: soft-real-time QoS sensitive, large data movement, and background data movement. The paper by Dong Lu and Peter Dinda addresses the important problem of generating synthetic, yet realistic Computational Grids for analysis and simulation purposes. The GridG generator is able to generate Grid network topologies that conform to well-known Internet power laws and is hierarchical in structure. The authors point out that more Grid data is needed to generate host resources and host/network correlations. The paper by Rich Wolski describes the Network Weather Service, a resource predictor tool that has been applied to networks and hosts in a Grid environment. It utilizes a mix-of-experts approach to adaptively select the best predictor using past values or postcasting. It has been shown to outperform single-point methods and has been widely used by Grid infrastructure and applications.

/lp/association-for-computing-machinery/guest-editorial-special-issue-on-grid-computing-luovWeJyKh